The use of 'nocapture' for parameters and return values is incorrect for C++ types, as they can actually capture a pointer into its own value (e.g. std::string in libstdc++)
rdar://115062687
This patch replace all in-memory objects of DebugValueAddrInst with
DebugValueInst + op_deref, and duplicates logics that handles
DebugValueAddrInst with the latter. All related check in the tests
have been updated as well.
Note that this patch neither remove the DebugValueAddrInst class nor
remove `debug_value_addr` syntax in the test inputs.
When we generate code that asks for complete metadata for a fully concrete specific type that
doesn't have trivial metadata access, like `(Int, String)` or `[String: [Any]]`,
generate a cache variable that points to a mangled name, and use a common accessor function
that turns that cache variable into a pointer to the instantiated metadata. This saves a bunch
of code size, and should have minimal runtime impact, since the demangling of any string only
has to happen once.
This mostly just works, though it exposed a couple of issues:
- Mangling a type ref including objc protocols didn't cause the objc protocol record to get
instantiated. Fixed as part of this patch.
- The runtime type demangler doesn't correctly handle retroactive conformances. If there are
multiple retroactive conformances in a process at runtime, then even though the mangled string
refers to a specific conformance, the runtime still just picks one without listening to the
mangler. This is left to fix later, rdar://problem/53828345.
There is some more follow-up work that we can do to further improve the gains:
- We could improve the runtime-provided entry points, adding versions that don't require size
to be cached, and which can handle arbitrary metadata requests. This would allow for mangled
names to also be used for incomplete metadata accesses and improve code size of some generic
type accessors. However, we'd only be able to take advantage of the new entry points in
OSes that ship a new runtime.
- We could choose to always symbolic reference all type references, which would generally reduce
the size of mangled strings, as well as make runtime demangling more efficient, since it wouldn't
need to hit the runtime caches. This would however require that we be able to handle symbolic
references across files in the MetadataReader in order to avoid regressing remote mirror
functionality.
This is essentially a long-belated follow-up to Arnold's #12606.
The key observation here is that the enum-tag-single-payload witnesses
are strictly more powerful than the XI witnesses: you can simulate
the XI witnesses by using an extra case count that's <= the XI count.
Of course the result is less efficient than the XI witnesses, but
that's less important than overall code size, and we can work on
fast-paths for that.
The extra inhabitant count is stored in a 32-bit field (always present)
following the ValueWitnessFlags, which now occupy a fixed 32 bits.
This inflates non-XI VWTs on 32-bit targets by a word, but the net effect
on XI VWTs is to shrink them by two words, which is likely to be the
more important change. Also, being able to access the XI count directly
should be a nice win.
It hass been a longstanding principle in LLVM that the presence of
debug info shall not affect code generation. This patch brings the
Swift frontend closer to this ideal:
- unconditionally emit shadow copies
- unconditionally bind type metadata
The extra allocas, bitcasts, geps, and stores being emitted get
optimized away when compiling at anything but -Onone. There are few
use-cases for compiling at -Onone without -g, so this shouldn't affect
performance for any real-world use-cases.
And update the existential container's initializeWithTake implementation
in the runtime. After only allowing bitwise takable values in the
inline buffer we can use memcpy to move existential container values.
rdar://31414907
SR-343
I was going to put this off for awhile, but it turns out that a lot of
my testcases are enums with multi-payload cases, which we currently
compile as tuples, so they were all still hanging until this patch.
I am going to leave in the infrastructure around this just in case. But there is
no reason to keep this in the tests themselves. I can always just revert this
and I don't think merge conflicts are likely due to previous work I did around
the tooling for this.
This includes global generic and non-generic global access
functions, protocol associated type access functions,
swift_getGenericMetadata, and generic type completion functions.
The main part of this change is that the functions now need to take
a MetadataRequest and return a MetadataResponse, which is capable
of expressing that the request can fail. The state of the returned
metadata is reported as an second, independent return value; this
allows the caller to easily check the possibility of failure without
having to mask it out from the returned metadata pointer, as well
as allowing it to be easily ignored.
Also, change metadata access functions to use swiftcc to ensure that
this return value is indeed returned in two separate registers.
Also, change protocol associated conformance access functions to use
swiftcc. This isn't really related, but for some reason it snuck in.
Since it's clearly the right thing to do, and since I really didn't
want to retroactively tease that back out from all the rest of the
test changes, I've left it in.
Also, change generic metadata access functions to either pass all
the generic arguments directly or pass them all indirectly. I don't
know how we ended up with the hybrid approach. I needed to change all
the code-generation and calls here anyway in order to pass the request
parameter, and I figured I might as well change the ABI to something
sensible.
Whan an alloca is dynamic, all FastISel can do is describe the
register in which the address of of the dynamic alloca is kept, and if
the value isn't used it will get kicked out by regalloc.
rdar://problem/36663932
* Remove RegisterPreservingCC. It was unused.
* Remove DefaultCC from the runtime. The distinction between C_CC and DefaultCC
was unused and inconsistently applied. Separate C_CC and DefaultCC are
still present in the compiler.
* Remove function pointer indirection from runtime functions except those
that are used by Instruments. The remaining Instruments interface is
expected to change later due to function pointer liability.
* Remove swift_rt_ wrappers. Function pointers are an ABI liability that we
don't want, and there are better ways to get nonlazy binding if we need it.
The fully custom wrappers were only needed for RegisterPreservingCC and
for optimizing the Instruments function pointers.
On architectures where the calling convention uses the same argument register as
return register this allows the argument register to be live through the calls.
We use LLVM's 'returned' attribute on the parameter to facilitate this.
We used to perform this optimization via an optimization pass. This was ripped
out some time ago around commit 955e4ed652.
By using LLVM's 'returned' attribute on swift_*retain, we get the same
optimization from the LLVM backend.
To make this stick, I've disallowed direct use of that overload of
CreateCall. I've left the Constant overloads available, but eventually
we might want to consider fixing those, too, just to get all of this
code out of the business of manually remembering to pass around
attributes and calling conventions.
The test changes reflect the fact that we weren't really setting
attributes consistently at all, in this case on value witnesses.
Use the generic type lowering algorithm described in
"docs/CallingConvention.rst#physical-lowering" to map from IRGen's explosion
type to the type expected by the ABI.
Change IRGen to use the swift calling convention (swiftcc) for native swift
functions.
Use the 'swiftself' attribute on self parameters and for closures contexts.
Use the 'swifterror' parameter for swift error parameters.
Change functions in the runtime that are called as native swift functions to use
the swift calling convention.
rdar://19978563
Allocate buffers for local generic/resilient values on the stack. alloc_stack
instructions in the entry block are translated using a dynamic alloca
instruction with variables size. All other alloc_stack instructions in addition
use llvm's stacksave/restore instrinsics to reset the stack (they could be
executed multiple times and with varying sizes).
Swift uses rt_swift_* functions to call the Swift runtime without using dyld's stubs. These functions are renamed to swift_rt_* to reduce namespace pollution.
rdar://28706212
The recent change to destructure tuples in SIL function return types
introduced some runtime changes where it was assumed that a SIL
function type like
$@convention(thin) () -> (@out X, @out Y)
Would have the same calling convention as the following C function:
void foo(void *X, void *Y);
Unfortunately, this only worked on x86-64, because the first @out
parameter in a SIL function type was lowered with the LLVM
'sret' attribute, which on i386 and ARM64 is not the same as the
first parameter to a function.
On i386, this manifested as a crash in a variety of executable tests
with a misaligned stack; on ARM64, a similar crash would occur because
the return value was initialized through the wrong register.
Hack around this by simply disabling 'sret' if a SIL function type
has multiple indirect return values.
Fixes <rdar://problem/24727411>.
Similarly to how we've always handled parameter types, we
now recursively expand tuples in result types and separately
determine a result convention for each result.
The most important code-generation change here is that
indirect results are now returned separately from each
other and from any direct results. It is generally far
better, when receiving an indirect result, to receive it
as an independent result; the caller is much more likely
to be able to directly receive the result in the address
they want to initialize, rather than having to receive it
in temporary memory and then copy parts of it into the
target.
The most important conceptual change here that clients and
producers of SIL must be aware of is the new distinction
between a SILFunctionType's *parameters* and its *argument
list*. The former is just the formal parameters, derived
purely from the parameter types of the original function;
indirect results are no longer in this list. The latter
includes the indirect result arguments; as always, all
the indirect results strictly precede the parameters.
Apply instructions and entry block arguments follow the
argument list, not the parameter list.
A relatively minor change is that there can now be multiple
direct results, each with its own result convention.
This is a minor change because I've chosen to leave
return instructions as taking a single operand and
apply instructions as producing a single result; when
the type describes multiple results, they are implicitly
bound up in a tuple. It might make sense to split these
up and allow e.g. return instructions to take a list
of operands; however, it's not clear what to do on the
caller side, and this would be a major change that can
be separated out from this already over-large patch.
Unsurprisingly, the most invasive changes here are in
SILGen; this requires substantial reworking of both call
emission and reabstraction. It also proved important
to switch several SILGen operations over to work with
RValue instead of ManagedValue, since otherwise they
would be forced to spuriously "implode" buffers.
